Mythra Code 1.20.1

Mythra Code 1.20.1

• Fix a production startup failure that could leave Mythra Code 1.20.0 showing a blank window.
• Keep the optimized startup bundle while correcting the dependency initialization order.
• Add production-renderer startup checks before packaging, including opening Settings and reloading the app.

This update preserves existing projects, chats, and settings.

If 1.20.0 opens a blank window, quit it and install 1.20.1 directly from the download for your platform. The blank interface cannot open its update controls. Do not delete the app's data folder.

Mythra Code 1.19.0

Mythra Code 1.19.0

Mythra Code 1.19.0 adds a local usage dashboard, clearer pricing evidence, and conversation and update-flow improvements.

Highlights

• Explore locally recorded usage in Settings with Overview, Models, and Compare views. See daily or weekly trends, token types, per-model breakdowns, per-prompt averages, and accessible data tables.
• Refresh and inspect pricing evidence from official OpenAI, Anthropic, and Cursor pages, Mythra's catalog, and OpenRouter. The app shows source status and pricing coverage, and distinguishes Claude's 5-minute and 1-hour cache writes. Supported historical estimates can be corrected when dated evidence supports the rate.
• Follow conversations more reliably: assistant-authored line breaks are preserved, streamed text and turn cleanup are steadier, and loading earlier transcript content better preserves your scroll position.
• Add feedback to selected Review diff lines more reliably, including in long diffs.
• Scroll long approval-question dialogs, and follow update download, install, and restart progress in a notice that does not shift the conversation.
• Inspect expanded performance scorecards. An opt-in native comparison harness can collect pilot process, window, and idle-resource measurements.

Notes

• Usage costs are local estimates at standard API rates, not subscription bills or provider invoices. OpenRouter reported charges stay separate. Detailed dated and per-model history starts with usage recorded after this upgrade; older aggregate usage cannot be reconstructed into that history.
• Pricing sources can be unavailable. Mythra shows each source's status and keeps validated fallback rates; it does not represent an unverified rate as current. Historical estimates change only when saved evidence supports a correction.
• Performance measurements describe the tested builds and environment; they do not establish a general speed improvement. The native harness's complete capture loop has not yet been validated on either operating system.
• Windows installers are signed for updates but not Authenticode-signed; Windows may show an Unknown publisher warning.

Mythra Code 1.18.2

Mythra Code 1.18.2

Mythra Code 1.18.2 improves prompt queues, review feedback, project checks, and reusable workflows.

Highlights

• Edit queued or failed prompts in place. Their order and attachments stay intact, your other draft is preserved, and delivery waits while an item is being edited—even after switching threads or reopening the app.
• Select part of an assistant reply or a Review diff to add anchored feedback. Edit notes and send them together, with or without a typed prompt.
• Save a check command for each project and run or stop checks explicitly from Review. When no command is saved, Find checks can inspect the project for an existing test command. A failed result can be added to feedback without sending it automatically.
• Add an optional setup command before each Run. Setup and launch share the same shell environment; a failed setup prevents launch.
• Create and share portable workflow recipes. Review imported files before saving, choose a project for manual runs, preview inputs and commands, and launch from Settings, Workspace, the command palette, or an explicitly selected ! recipe in the composer.
• Run saved agent workflows through Claude Code and Cursor sessions across steps. Workflow stop, checkpoint, project routing, transcript resume, and quoted Windows shell path handling are more reliable.

Notes

• Project checks and imported recipes do not run automatically. Review a recipe before saving or sharing it: text you put inside its prompts, commands, and inputs is included in the export.
• Claude Code and Cursor support agent steps in saved workflows; explicit shell-command steps still require the Codex runtime.
• Windows installers are signed for updates but not Authenticode-signed; Windows may show an Unknown publisher warning.

Mythra Code 1.18.1

Mythra Code 1.18.1

Mythra Code 1.18.1 improves first-run setup, conversation titles, and provider controls, with a visual refresh across the app.

Highlights

• Explore a shorter five-page onboarding tour with provider-specific setup guidance, optional project and Git details, and previews for themes, fonts, and effort-slider styles.
• Carry provider and appearance choices into Settings as unsaved drafts. Save applies them; Cancel restores saved choices and returns to the same tour page.
• See automatic conversation titles appear with a brief reveal, and find title-generation controls in Settings → Projects.
• See signed-out OpenAI and Claude choices clearly marked as unavailable. Selecting one explains where to sign in while preserving saved choices.
• Configure sub-agents with one main switch, a parallel limit, and explicit provider and model choices. Collapsed worker cards take less space.
• Find calmer surfaces and outlined actions throughout the interface, plus refreshed Sol and Luna model artwork.
• Fix Claude Code configuration for title generation and run-command discovery.

Notes

• Automatic title generation remains opt-in and makes a separate provider request when enabled, which may incur provider cost.
• Theme, font, and slider previews do not change saved preferences until applied in Settings.
• Windows installers are signed for updates but not Authenticode-signed; Windows may show an Unknown publisher warning.

Mythra Code 1.18.0

Mythra Code 1.18.0

Mythra Code 1.18.0 adds local Git controls and a GitHub pull-request workflow for project threads.

Highlights

• Create and switch local branches, inspect changes, stage or unstage selected files, and commit from the Git panel.
• Attach an existing GitHub pull request to a thread or create a new draft or ready pull request. View its status, checks, and review state from the thread.
• Merge an attached pull request using a method allowed by its repository, then explicitly update a clean local project branch after the GitHub merge.
• Optionally publish new branches and commits automatically for a project. This setting is off by default and resumes pending committed work when Mythra Code next opens the project.
• See attached pull requests in the thread inbox, optionally archive a thread after its merge, and generate conversation titles with a selected provider and model.
• Refresh the OpenAI model catalog after a Codex runtime update so newly available models can appear without replacing the selected model or interrupting active work.
• Show the Sol and Luna artwork for GPT-6 Sol and GPT-6 Luna in the model picker.

Notes

• Local Git controls work without GitHub. Automatic publishing only pushes committed work; it never stages, commits, or changes local files on its own.
• The pull-request workflow supports github.com and one primary pull request per thread. New pull requests use the configured repository; existing fork pull requests can be attached by URL.
• Automatic title generation is off by default and makes a separate provider request when enabled, which may incur provider cost.
• If a Codex task, terminal command, workflow command, or approval is active when the installed Codex runtime changes, OpenAI model refresh keeps the existing catalog and asks you to retry after that work finishes.
• Automated checks cover the relevant Git, pull-request, thread, and native state transitions. Actual GitHub behavior can still depend on repository rules, credentials, and network conditions.
• Windows installers are updater-signed but not Authenticode-signed; Windows may show an Unknown publisher warning.

Mythra Code 1.17.5

Mythra Code 1.17.5

Mythra Code 1.17.5 improves transcript durability and keeps background terminal, usage, health, Skills, and menu work attached to the correct project and lifecycle.

Highlights

• Preserve short turns that complete inside the local transcript save debounce instead of treating the save as metadata-only.
• Persist every intervening completed turn when multiple turns finish between saves, while retaining generation checks and safe retry behavior.
• Keep long-running terminal processes routed to the project that launched them even after many later commands complete.
• Preserve the terminal scrollback currently on screen while other projects create additional idle terminal sessions.
• Coalesce a forced provider-usage refresh requested during an in-flight read into one follow-up request instead of silently dropping it.
• Prevent delayed thread-health results from closing a turn after that turn has reported newer progress.
• Let slow Skills scans finish before the polling watcher starts another scan, while explicit edits and focus refreshes can still supersede older work.
• Coalesce select-menu scroll, resize, and observer notifications into one repositioning measurement per animation frame and avoid unchanged state updates.

Notes

• Existing transcript storage remains local-first. This release adds no new network path, telemetry, raw transcript diagnostic logging, provider setting, or user-data migration.
• The durability changes apply to paged local Claude and Cursor transcripts and retain the existing full-history snapshot behavior.
• Animations, streaming cadence, styling, and background completion recovery are unchanged.
• The additional correctness guards add about 0.19% to total JavaScript size. The checked performance limits were updated to the measured builds; CSS and dependency counts are unchanged.
• Automated verification covers the affected persistence, terminal, usage-refresh, thread-health, Skills, and real-browser menu behavior. Paid live-provider smoke tests remain opt-in.
• Windows installers are updater-signed but not Authenticode-signed; Windows may show an Unknown publisher warning.

Mythra Code 1.17.4

Mythra Code 1.17.4

Mythra Code 1.17.4 keeps run-command discovery working while you move around the app, rejects provider progress narration before it can become a Run command, and records safer diagnostics when a native discovery fails.

Highlights

• Keep each discovery worker in a project-keyed background store so switching projects, opening Chats, or collapsing the project control no longer cancels the investigation.
• Save a successful result back to the project that started the discovery, even when another project is currently selected, and restore its in-progress or saved state when you return.
• Preserve explicit Stop behavior: cancellation still suppresses late results and never overwrites a project after the user stops discovery.
• Ask native OpenAI, Claude, and Cursor workers to investigate silently and return one final JSON result instead of schema-shaped progress updates.
• Reject observed narration patterns such as status updates, first-person investigation messages, and project-description sentences before they can be saved as executable commands.
• Keep valid launchers and shell commands—including English-named tools and builtins such as next, check, inspect, read, and let—instead of broadly rejecting sentence-like command text.
• Prefer the most relevant transient provider error line over echoed tool output, making authentication and process failures easier to understand in the discovery popover.
• Write only sanitized provider/model/effort/project metadata, a fixed failure category, and stdout/stderr byte counts to run-discovery-last-failure.log. Raw provider output, error text, source excerpts, and credentials are never persisted there.

Notes

• Discovery remains explicit and advisory. It saves a proposed command but never launches the application; the user still starts it with Run.
• Native discovery workers remain bounded by the existing concurrency limit, five-minute timeout, output limits, process-tree cancellation, temporary-session cleanup, and provider sandbox or tool controls.
• Native provider tools rely on those provider controls and discovery instructions rather than Mythra Code's app-owned project-path filter. Use native discovery only with trusted projects and review the saved command before running it.
• OpenRouter and LM Studio continue to use Mythra Code's bounded, path-confined inspection operations and retained multi-round context.
• No provider conversation or raw provider output is retained by the new failure log. The latest fixed-category diagnostic is stored only in the local app-data folder to help distinguish authentication, permission, timeout, invalid-response, process-exit, cancellation, and provider failures.
• The release adds no dependencies or permanent polling.
• Windows installers are updater-signed but not Authenticode-signed; Windows may show an Unknown publisher warning.

Mythra Code 1.17.3

Mythra Code 1.17.3

Mythra Code 1.17.3 makes run-command discovery investigate projects more like a focused native coding task, then saves the verified development command directly to the project’s Run button without launching it.

Highlights

• Give OpenAI, Claude, and Cursor one continuous temporary inspection session rooted in the selected project, with provider-specific read-only or ask-mode safety controls and isolated session storage.
• Preserve the inspection conversation for OpenRouter and LM Studio so those providers can gather evidence across multiple directory, file, search, and executable-location requests instead of restarting from a fresh prompt each round.
• Detect project-local Python environments such as .venv, venv, and env, inspect their configuration and installed-package metadata, and prefer the interpreter that already contains the project’s dependencies.
• Save a successful discovery result directly to Run through the normal project-setting update. Discovery never launches the application; the user still starts it explicitly with the Run button.
• Cancel cleanly on Stop or project navigation and suppress late results so an old discovery cannot overwrite either the previous or newly selected project.
• Disable conflicting manual Run-command edits while discovery is active, then restore editing after completion or cancellation.
• Remove only a redundant cd into the selected project from final commands, while preserving meaningful directory changes into nested applications or packages.
• Default new discovery preferences to OpenAI GPT-5.6 Luna with High reasoning and Fast service while preserving existing saved provider, model, and effort choices.

Notes

• Discovery is explicit and advisory. Workers are instructed not to launch applications, run project scripts, install dependencies, or modify project files.
• Native OpenAI, Claude, and Cursor investigation is bounded by a five-minute timeout, output limits, process-tree cancellation, temporary-session cleanup, and each provider’s sandbox or tool controls.
• Native provider tools rely on those provider controls and discovery instructions rather than Mythra’s app-owned project-path filter. Use native discovery only with trusted projects and review the saved command before running it.
• OpenRouter and LM Studio retain bounded inspection history for up to twelve inspection rounds plus a final response. Their app-owned file operations remain path-confined and exclude generated, sensitive, binary, and symbolic-link file reads.
• Bounded project source and configuration evidence is sent to the provider selected for discovery. Sensitive filename and line filtering is best-effort, so do not use discovery on projects that embed secrets under innocuous names.
• Failed temporary cleanup is surfaced and retained for a later cleanup attempt instead of being silently discarded.
• The release adds no dependencies or permanent polling. Verified startup and total JavaScript sizes decreased slightly, with existing performance budgets unchanged.
• Windows installers are updater-signed but not Authenticode-signed; Windows may show an Unknown publisher warning.

Mythra Code 1.17.2

Mythra Code 1.17.2

Mythra Code 1.17.2 lets run-command discovery investigate real project evidence instead of depending on a README, manifest, or predefined run script.

Highlights

• Infer development commands from source entry points, imports, framework or engine configuration, dependencies, scripts, folder structure, and other project evidence.
• Use the same app-owned inspection protocol with OpenAI, Claude, Cursor, OpenRouter, and LM Studio, including providers without native tool support.
• Let the selected model request bounded relative directory listings, text-file reads, literal searches, and installed-executable location checks without executing candidate programs.
• Preserve documentation text, blank lines, and command blocks so uncommon launch commands—such as application-bundle executables—are not removed by ecosystem keyword filtering.
• Discover commands for undocumented projects and monorepos through multiple evidence rounds while keeping the final suggestion reviewable before the user explicitly uses or saves it.
• Keep project inspection read-only, cancellable, isolated from conversation history, and confined to the selected project; generated folders, sensitive filenames, out-of-tree paths, binaries, and symbolic links are excluded.

Notes

• Discovery is advisory and opt-in. It never edits project files, installs dependencies, executes the proposed command, automatically saves it, or creates a visible conversation.
• Investigation is bounded to four inspection rounds plus a final model response, with per-operation traversal, file, output, timeout, and accumulated-evidence limits.
• Bounded project source and configuration excerpts are sent to the provider selected for discovery. Sensitive filename and line filtering is best-effort, so do not use discovery on projects that embed secrets under innocuous names.
• Installed-executable checks inspect filesystem metadata only; they do not launch the executable.
• The release adds no dependency, frontend bundle growth, permanent polling, or unsolicited provider request and includes expanded native protocol, provider-adapter, path-confinement, cancellation, and limit coverage.
• Windows installers are updater-signed but not Authenticode-signed; Windows may show an Unknown publisher warning.

Mythra Code 1.17.1

Mythra Code 1.17.1

Mythra Code 1.17.1 improves transcript and image-attachment reliability and adds an explicit, reviewable way to discover project run commands.

Highlights

• Preserve local Claude and Cursor history when renaming conversations that have not yet been opened, including legacy transcript migration and serialized metadata updates.
• Keep multi-image pastes ordered and durable across asynchronous preparation, Send, Steer, and project navigation without delivering a prepared draft to the wrong task.
• Reject unsupported HEIC/HEIF images with conversion guidance instead of mislabeling them, and use browser-native encoding with a 50 MB pre-encoding limit.
• Bound failed transcript-save retries, retain retryable dirty state, and offer an explicit Keep open or Close without saving decision when the close-time flush cannot finish safely.
• Preserve durable timeline ordering when provider output rehydrates and prevent completed asynchronous sends from stealing the currently selected task.
• Find a suggested run command from the Run popover using OpenAI, Claude, Cursor, OpenRouter, or LM Studio, then explicitly review, use, or save it; discovery never starts the command automatically.
• Keep discovery isolated from conversations and local transcripts, use bounded metadata-only project context, remember provider-specific model settings, and support cancellation and cleanup when switching projects or closing the app.

Notes

• Run-command discovery defaults to OpenAI GPT-5.6 Luna with Low reasoning and Fast service, while remaining configurable per provider.
• Discovery is advisory and opt-in. It creates no visible task, performs no permanent polling, and does not automatically save or execute a proposed command.
• Automatic deletion of durable image storage remains deferred until Mythra Code can prove that no transcript, draft, queue, legacy file, or provider history references an image.
• Existing legacy external-image previews remain supported; tighter authorization requires a future reference-aware migration to avoid breaking saved histories.
• The release adds no dependency and includes expanded frontend, Chromium, WebKit, Rust, persistence, attachment, provider-discovery, cancellation, and native lifecycle coverage.
• Windows installers are updater-signed but not Authenticode-signed; Windows may show an Unknown publisher warning.